EP0139446B1 - Anlage zum erkennen unbekannter MusterE - Google Patents

Anlage zum erkennen unbekannter MusterE Download PDF

Info

Publication number
EP0139446B1
EP0139446B1 EP84306098A EP84306098A EP0139446B1 EP 0139446 B1 EP0139446 B1 EP 0139446B1 EP 84306098 A EP84306098 A EP 84306098A EP 84306098 A EP84306098 A EP 84306098A EP 0139446 B1 EP0139446 B1 EP 0139446B1
Authority
EP
European Patent Office
Prior art keywords
pattern
matrix
patterns
memory
mask
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Expired - Lifetime
Application number
EP84306098A
Other languages
English (en)
French (fr)
Other versions
EP0139446A3 (en
EP0139446A2 (de
Inventor
Kenichi Patent Division Toshiba Corp. Maeda
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Toshiba Corp
Original Assignee
Toshiba Corp
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Toshiba Corp filed Critical Toshiba Corp
Publication of EP0139446A2 publication Critical patent/EP0139446A2/de
Publication of EP0139446A3 publication Critical patent/EP0139446A3/en
Application granted granted Critical
Publication of EP0139446B1 publication Critical patent/EP0139446B1/de
Anticipated expiration legal-status Critical
Expired - Lifetime legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning

Definitions

  • the present invention relates to the field of pattern recognition apparatus and, more particularly, is directed to a pattern recognition apparatus which can recognise malformed and distorted patterns.
  • Pattern recognition apparatus is commonly used to convert information patterns, such as speech and written characters, to a form suitable for processing and storage by computer systems.
  • Conventional pattern recognition methods are usually based on either a pattern matching scheme or a character extraction scheme.
  • pattern matching schemes the identity of a given input pattern is established according to the degree of its similarity to a specific reference pattern. Similarity methods are among the most effective pattern recognition techniques based on a pattern matching scheme. In these methods, a similarity value if calculated between an input pattern and one or more reference patterns for each category of a given pattern set. A category includes all the expected variations of a particular pattern. A comparison between similarity values is used to determine the category to which the input pattern belongs.
  • the simple similarity method is the most basic of the similarity methods. As shown in Figure 1A, this method utilizes vectors to represent an input and a reference pattern. The similarity value is calculated as an angle 8 between an input pattern vector A and a reference pattern vector B. The closer angle ⁇ is to 0, the closerthe input pattern matches the reference pattern. Thus, when angle 8 is 0, the input pattern is identical to the reference pattern. In actual practice, an input pattern is identified as belonging to a particular category when angle ⁇ is within a predetermined range.
  • the similarity value S which exist between an input pattern f and a reference pattern fo may be expressed as:
  • Any pattern drawn on a plane can be expressed by an n-dimensional vector in the following manner.
  • the plane is divided into n picture elements or cells, each of which has a darkness or density ranging from white through gray to black as a function of its position on the plane. If the positions of the picture elements are expressed as x 1 , x 2 ,..., x n , and the darkness of the picture elements are expressed as f(x 1 ), f(x 2 ),..., f(x n ), respectively, vector f can be uniquely defined in n-dimensional coordinates where f(x 1 ), (fx 2 ), ..., f)x n ) correspond to the projections of vector f on the coordinate axis 1, 2,..., n.
  • Equation (1) The simple similarity defined by equation (1) means that S[f,f o ] takes maximum value 1 when vectors f and f o in n-dimensional coordinates are parallel and that S[f,f o ] takes minimum value 0 when the vectors are perpendicular. Thus S[f,f o] varies from value 1 where two patterns on the plane are overlapped, to value 0 where two patterns are quite different from each other.
  • the simple similarity method has the advantage that the design of a dictionary of standard reference patterns can be created rather easily. This method is also not greatly affected by such local noise as stains or scratches in the patterns. The simple similarity method is affected adversely, however, by such overall changes in the patterns as occur in handwritten letters or voice sound patterns.
  • tpm 1, 2 ... M
  • S X the primary and secondary reference patterns.
  • the value for S X. also varies from 0 to 1 in accordance with similarity between the pattern f and the set of reference patterns tpm.
  • the similarity value is calculated as an angle 8 between vector A, representing an input pattern, and subspace C constructed from a plurality of vectors representing a particular category of patterns.
  • Subspace C is usually in m-dimensional space but is illustrated in Figure 1B as a two-dimensional plane.
  • Each vector is associated with a deformation or variation of primary reference pattern B so that a range of malformed or distorted input patterns may be correctly recognized.
  • the multiple similarity method is useful for recognition of patterns which are subject to overall deformations, the range of deformation determination is limited. That is, only deformations representing small variations of the input patterns can be detected. With respect to handwritten characters and speech, however, there are many more possible variations for the same pattern as compared to printed characters. Because the multiple similarity method can only include a limited number of these variations, the method is not reliable for recognizing written characters and speech patterns which can have a multitude of pattern variations.
  • One such method is the combination of the multiple similarity method and another recognition method such as one based on a character extraction scheme.
  • Another recognition method such as one based on a character extraction scheme.
  • An example of a pattern recognition method based on a character extraction scheme is disclosed in U.S. Patent No. 3,541,511.
  • this method various kinds of characters are extracted from various parts of the patterns and the unknown patterns are recognised by a combination of reference character patterns.
  • the problem in this method has been that complicated algorithmic processing is required. Moreover, a large amount of memory is required for storing the reference patterns. There is also the difficulty that this method cannot be easily automated and much labor is needed to use it.
  • a method combining the multiple similarity method and the character extraction method is disclosed in the preliminary report for the National Convention of the Electrocommunication Society, page 5-378, published by the Japan Electrocommunication Society. According to this method, the multiple similarity method is applied to the unknown patterns as a first step. In a second step, the character extraction method is applied to the result. While this combined method is useful for the recognition of complicated handwritten letters, the apparatus for realising the method is, however, complicated and automatic design of the reference pattern dictionary is also difficult.
  • the present invention seeks to provide a new and improved pattern recognition apparatus which is capable of extending the deformation range in which patterns may be recognised.
  • the present invention also seeks to provide a new and improved pattern recognition apparatus which is capable of extending the deformation range in which patterns may be recognised without increasing memory storage requirements over that required by conventional pattern recognition apparatus.
  • a pattern recognition apparatus for determining the category to which an unknown input pattern belongs, comprising: a memory for storing a plurality of reference patterns for each category of a predetermined set of reference pattern recognitions, which reference patterns are orthonormalised with respect to each other, and input vector generating means for generating an input vector representing features from the unknown input pattern; characterised by a basis generating means coupled to the input vector generating means for deforming the input vector to produce a plurality of vectors, and for generating base vectors which are orthonoarmalised with respect to each other; matrix producing means coupled to the memory and the basis generating means for producing a matrix for each category of the said sets of reference patterns, each said matrix being formed from a plurality of data elements derived from the base vectors generated by basis generating means and from the reference patterns stored in said memory, the matrix producing means including projector generating means for generating a first projector corresponding to the bases generated by the basis generating means and a second projector corresponding to said reference patterns,
  • FIG. 2 illustrates in simplified form the principle concept of the present invention.
  • reference patterns are represented by subspace C.
  • Subspace C is constructed from a plurality of vectors representing a particular category of patterns.
  • Subspace C is usually in M-dimensional space but is illustrated in Figure 2 in a two-dimensional plane.
  • Each vector is associated with a deformation or variation of primary reference pattern B.
  • the input pattern is represented by subspace D which is constructed from a plurality of vectors.
  • Subspace D is also usually in M-dimensional space but is illustrated in Figure 2 in a two-dimensional plane.
  • Each vector forming subspace D is associated with a deformation or variation of primary input pattern A.
  • Angle 6 between subspace C and subspace D is used as the similarity value.
  • Angle 8 is defined as a minimum angle between one of the vectors included in subspace C and one of the vectors included in subspace D.
  • FIG. 3 is a flow chart showing the overall operation of the present invention.
  • An unknown input pattern f is provided in step 1.
  • bases q l o , ⁇ 1 and ⁇ 2 are examples of bases q l o , ⁇ 1 and ⁇ 2 .
  • step 3 angle e between the input subspace ( ⁇ 0 , ⁇ 1 , ⁇ 2 ) and a second subspace constructed from a plurality of vectors representing a reference pattern for a certain category is calculated. In practice, it is convenient to calculate values that relate to angle 8, for example cos 2 S. These values are calculated for every category.
  • step 4 a decision is made to determine the category to which input pattern f belongs by comparing the values calculated in step 3.
  • Figure 4 shows a detailed flow chart of step 2 in Figure 3.
  • input pattern f is normalized to adjust its size to the size of the reference patterns.
  • the normalized input pattern f is then blurred in step 2.
  • blurring is used for reducing the noise components of the input pattern.
  • the blurring operation is executed by the following equation: where f indicates the normalized input pattern, g indicates a blurred pattern and G is a Gaussian function.
  • the integral calculation included in equation (3) is replaced by a masking operation according to the present invention.
  • step 3 the amount of shift of blurred pattern g is approximated and a differential value 9 x to the x-direction and a differential value 9 y to the y-direction are calculated as follows: These partial differential operations are also performed by a masking operation in the present invention.
  • step 4 pattern g, gx and gy are orthonormalized to create bases ( ⁇ 0 , ⁇ 1 , ⁇ 2 ) for the input subspace as follows:
  • ⁇ g ⁇ indicates a norm or magnitude of the pattern (vector) g.
  • Figure 5 shows a flow chart for calculating angle ⁇ between the input subspace and reference patterns.
  • Angle 8 between subspaces U and V is defined as follows Equation (8) shows that the angle formed by subspaces U and V is defined as the smallest angle of many angles formed between one vector of subspace U and another vector of subspace V.
  • the calculation of equation (8) is obtained by utilizing projectors P and Q of subspaces U and V.
  • Projectors P and Q are calculated as follows:
  • ⁇ > indicates a dyad operation.
  • vector u can be resolved as:
  • This value represents the maximum eigenvalue of matrix PQP and is equal to the maximum eigenvalue of matrix QPQ. Therefore, it is possible to calculate cos 2 ⁇ between subspaces U and V by obtaining the maximum eigenvalue of matrix PQP or QPQ. For the purpose of obtaining eigenvalues, a power method may be used.
  • total space is a three-dimensional space whose coordinate system 8 consist of x, y and z, unit vectors i, j and k on the x-axis, y-axis and z-axis are indicated by:
  • subspace U is along the x-y plane of the three-dimensional space
  • subspace V is along the z-axis.
  • projectors P and Q are as follows:
  • the maximum eigenvalue is determined as follows: In this case, the maximum eigenvalue becomes "0" so that cos 2 ⁇ is 0. This means that angle ⁇ between subspaces U and V is 90°.
  • projectors P and Q are as follows:
  • matrix PQP becomes:
  • the characteristic equation for matrix PQP is: Therefore, The maximum eigenvalue is "1" in this example so that angle 6 is 0°.
  • Figure 6 shows another flow chart for calculating angle 8 between the input subspace and a reference subspace in which the required number of calculation steps are reduced in comparison to the flow chart shown in Figure 5.
  • the eigenvalue relating to matrix QPQ can be considered as: Because vector v belongs to subspace V (v ⁇ V), it is expressed as follows: Thus, the left side of equation (11) is expressed as: The right side of equation (11) is expressed as: Therefore, ⁇ k in these equations: is satisfied. By expressing a m as: The following relationship may be used to produce the matrix rather than the relationship shown by equation (11) above: Accordingly, matrix X (equation 12) is generated in step 1 in Figure 6. The maximum eigenvalue of matrix X is calculated in step 2. The dimension of matrix X is small compared to the dimension of matrix QPQ.
  • Figure 7 is a flow chart of another example for obtaining bases.
  • the difference between Figures 4 and 7 is in the masking operation.
  • three mask patterns are prepared in advance.
  • the first mask pattern is that of the Graussian function as in Figure 4.
  • the second mask pattern is obtained by x-directional partial-differentiation of the Gaussian function and the third mask pattern is obtained by y-directional partial-differentiation of the Gaussian function.
  • three vectors are given by:
  • the orthornormalizing step 3 is the same as step 4 in Figure 4.
  • an unknown input pattern is expressed by the input subspace being n-dimensional space and reference patterns are expressed by the reference subspace being m-dimensional space.
  • the angle between the input subspace and the reference subspace is calculated as a measure of similarity.
  • FIG. 8 is a block diagram of a preferred embodiment of the pattern recognition apparatus according to the invention.
  • the apparatus can serve as a handwritten character reader which recognizes character patterns such as Japanese kanji-characters, Japanese kama-characters and alphanumeric characteristic.
  • Character patterns are written on a sheet and the sheet is divided into n picture elements or cells. Each cell has a darkness or density ranging from white to black as a function of its position on the sheet.
  • the sheet is optically scanned by a scanner in block 1 which usually comprises CCD (chared Coupled Device) sensors for obtaining electrical signals which represent the darkness or density of each picture element on the sheet.
  • the output signals of scanner 1 are converted to digital signals by an A/D (Analog to Digital) converter in block 2.
  • the digital signals obtained from the A/D converter are supplied to a pre-processing unit in block 3.
  • the pre-processing is used to eliminate noise and to normalize the size of the input character pattern to adjust it to the size of the reference patterns hereinafter described.
  • the normalized input pattern data in the form of digital signals associated with each of the picture elements on the sheet is stored in an input pattern memory in Block 4.
  • the input pattern data stored in the input pattern memory is hereinafter considered as being an input vector f where the position of the picture elements on the sheet are expressed as x" x 2 , ..., x" and the darkness of the picture elements are expressed as f(x i ) f(x 2 ), ..., f(x n ) ' respectively.
  • vector f may be uniquely defined in n-dimensional coordinates where f(x 1 ), f(x 2 ), ..., f(x n ) corresponod to the projections of vector f on the coordinate axis 1, 2, ..., n.
  • a basis generating unit is provided in block 5 for generating three bases ⁇ 0 , ⁇ 1 , and ⁇ 2 relating to input vector f.
  • the functions of the basis generating unit are:
  • a matrix generating unit is provided in block 7 for generating a QPQ matrix.
  • the matrix unit generates a projection operator Q from the three bases vectors, and a projection operator P from the vectors representing the reference patterns which are stored in reference pattern memory 8.
  • the matrix unit then calculates the QPQ matrix and supplies it to a matrix memory in block 9.
  • the reference patterns stored in the reference pattern memory in block 8 are the same type of patterns disclosed in U.S. Patent No. 3,906,446 and include a plurality of reference patterns for each category of pattern, each respective pattern being orthonormalized to the other patterns.
  • a reference pattern is indicated hereinafter by a vector ⁇ k m being the m-th reference pattern of the k-th category.
  • An eigenvalue calculating unit is provided in block 10 for obtaining eigenvalues for the QPQ matrix QPQ.
  • Eigenvalues are stored in an eigenvalue memory in block 11 and supplied to a decision unit in block 12.
  • the decision unit determines the maximum eigenvalue and identifies the category to which the input character pattern belongs.
  • FIG 9 shows an example of the basis generating unit in block 5 in Figure 8.
  • An address control circuit 100 supplies required addresses to the various memory units and selection signals to the various selector circuits in the basis generating unit. Lines for supplying selection signals are shown only by arrows to simplify the drawing.
  • a signal SO is received from the pre-processing unit in block 3 indicating the start of the operation.
  • Circuit 100 provides signal A to selector 101 for selecting terminal 1 of selector 101.
  • Circuit 100 also provides signal B to selector 102 and selector 103 for selecting their respective terminal 1.
  • Circuit 100 then successively supplies addresses to lines 104, 105, and 106, respectively.
  • the address on line 104 is supplied to the input pattern memory in block 4 so that it outputs the data for each of the picture elements of input vector f.
  • the address on line 105 is supplied to mask memories 107, 108 and 109 so that they output each data element for corresponding mask data.
  • the address on line 106 is supplied to blur pattern memory 110 and differential pattern memories 111, 112. Therefore, the data associated with each picture element of input vector f and each picture element of mask memory 107 are supplied to sum-of-product circuit 113 to produce a blurred vector g.
  • Circuit 113 executes a calculation between vector f and the mask data in a well-known manner such as by calculating the sum of the inner product of the vector f and the mask data. Accordingly, blurring of the input pattern is realized by the blurring mask data stored in mask memory 107.
  • a Gaussian function such as shown by the table in Figure 13A, is suitable for the blurring mask data.
  • the outputs of sum-of-product circuit 13 is successively stored in blurred pattern memory 110 through selector 103.
  • address control circuit 100 provides signal A indicating the selection of terminal 2 of selector 101 and signal B indicating the selection of terminal 2 of selectors 102 and 103.
  • the next operation is to obtain a differentiated vector gx.
  • mask memory 108 contains x-direction differential mask data such as shown by the table in Figure 13B.
  • Each data element of blurred vector g stored in blurred pattern memory 110 is successively read out according to the address on line 106 and is supplied to the sum-of-product circuit 113 through selector 101.
  • each data element of the x-direction differential mask data is supplied from mask memory 108 through selector 102.
  • the sum-of-product circuit 113 calculates each data element of partially differentiated vector gx and provides it to differential pattern memory 111 through selector 103.
  • address control circuit 100 After obtaining differentiated vector gx, address control circuit 100 outputs signal B indicating selection of terminal 3 of selectros 102 and 103 for the operation of y-directional partial differentiation of blurred vector g.
  • signal B indicating selection of terminal 3 of selectros 102 and 103 for the operation of y-directional partial differentiation of blurred vector g.
  • partial differentiated vector gy is obtained using y-direction differential mask data such as shown in the table in Figure 13C which is written into differential pattern memory 112.
  • Basis ⁇ 0 is obtained by normalizing blurred vector g stored in blurred pattern memory 110.
  • Address control circuit 100 provides selection signal C for indicating the selection of input terminal 1 of selector 120 and a selection signal D for indicating selection output terminal 1 of selector 120.
  • Circuit 100 then supplies addresses to blurred pattern memory 110.
  • Each data element of blurred vector g is successively supplied to norm circuit 121 through selector 120.
  • Norm circuit 121 calculates the norm or magnitude of the input vector and comprises a squaring circuit for squaring each data element, a summation circuit for adding the squared values and a square root circuit for taking the square root of the summed values.
  • Norm circuit 121 provides norm ⁇ g ⁇ of blurred vector to buffer 122.
  • Address control circuit 100 then provides selection signal D indicating the selection of output terminal 2 of selector 120. Circuit 100 also provides selection signal E indicating the selection of teminal 5 of selector 125 and then provides addresses to lines 106 and 126. Accordingly, each data element of blurred vector g is supplied to divider 124 through selector 120. Another input of divider 124 is supplied with data from the output of buffer 122. Divider 124 divides each data element of blurred vector g by norm
  • vector gx and gy are normalized. This operation is similar to the operation discussed above with respect to first base vector ⁇ 0 . That is, input terminal 2 and output terminal 1 of selector 120 are selected and supplies each data element of vector gx stored in differential pattern memory 111 to norm circuit 121. The calculated norm value is stored in buffer memory 122. Next, each data element of vector gx is supplied to divider 124 by selecting output terminal 2 of selector 120. The output of divider 124 is written into a first work or scratch pad memory 129 by selecting terminal 1 of selector 125. Work memory 129 stores normalized vector gx /
  • Address control circuit 100 then provides selection signal C indicating the selection of input terminal 4 and selection signal D indicating the selection of output terminal 1 of selector 120.
  • Work memories 129 and 130 successively output each data element of the normalized vectors according to addresses supplied from line 126.
  • Adder 131 adds these data elements and subtractor 132 subtract the data elements of work memory 130 from the data elements of work memory 129.
  • Input terminal 4 of selector 120 is then selected so that the output of adder 131 is supplied to norm circuit 121.
  • + gy / ⁇ g y ⁇ ⁇ is stored in memory buffer 122.
  • address control circuit 100 After the calculation of the norm value, address control circuit 100 provides selection signal D indicating the selection of output terminal 2 and supplies addresses to line 126 with selection signal E indicating the selection of terminal 4 of selector 125.
  • the output of adder 131 is supplied to divider 124 and the output of divider 124 is supplied to second memory 127.
  • Second memory 127 stores orthonormalized base ⁇ 1 .
  • third base ⁇ 2 is generated and stored in a third memory 128 using the output of subtractor 132.
  • Figure 10 shows an example of the matrix generating unit of block 7 in Figure 8.
  • Address control circuit 200 is operated in response to an end signal supplied from address control circuit 100 in Figure 9 at the end of the generation of the three base vectors.
  • Circuit 200 clears projection memories 201, 202.
  • the matrix generating unit first generates a projector Q.
  • Circuit 200 provides selection signals F1 and F2 to selector 203, each indicating the selection of input terminal 1 and output terminal 2 of selector 203, respectively.
  • Circuit 200 then supplies addresses to first base memory 126 and work memory 204 through lines 210 and 211 so that every data element of first base ⁇ 0 are transformed to work memory 204 through selector 203.
  • circuit 200 supplies selection signal F2, indicating the selection of output terminal 1 of selector 203, and selection signal G, indicating the selection of terminal 1 of selectors 207 and 208.
  • First base memory 126 outputs a data elemernt ⁇ 0j in response to an address j from circuit 200.
  • circuit 206 selects input terminal 2 and output terminal 2 and transfers second base ⁇ 1 into work memory 204. Thenn circuit 200 controls the address of memories 127, 204 and 201, so that the multiplied value ⁇ 1 j ⁇ 1 i is accumulated into the (i, j) data elements of projector Q of projector memory 201. Referring to the third basis ⁇ 2 , each data element is transferred to work memory 204. Each multiplied value ⁇ 2 j ⁇ 2 i is then accumulated into the (i, j) element data of projector Q. Therefore, projector memory 201 is now storing projector Q corresponding to three bases ⁇ 0 , ⁇ 1 and ⁇ 2 .
  • the second stage of operation of the matrix generating unit is to generate a projector P from the reference pattern which is similar to the first stage of operation except that circuit 200 accesses reference pattern memory 8 and projection memory 202.
  • Pk hereafter indicates projector P of the k-th category.
  • Each element data of the projector Pk is given by: wherein M is the number of reference patterns of k-th category.
  • circuit 200 supplies addresses to the memories, so that the output of multiplier 205 is accumulated into projector memory 202. Such operation is repeated m times for every reference pattern of the k-th category.
  • the k-th projector Pk corresponding to the k-th category is created in projector memory 202.
  • the third step of matrix generating unit 7 generates a multiplied matrix QPkQ using projectors Q and Pk.
  • Circuit 200 supplies a driving signal H to matrix multiplier 209 which multiplies projector Pk by the projector Q and thereafter multiplies projector Q by the multiplied matrix QPk.
  • the result is written into matrix memory 9 according to addresses supplied from circuit 100 through line 214.
  • Figure 11 shows an example of eigenvalue calculating unit 10 in Figure 8.
  • Unit 10 finds the maximum eigenvalue of the matrix in the matrix memory 9 according to the power method.
  • maximum detector 306 searches for the maximum eigenvalue ⁇ M from buffer 305.
  • the eigenvalue is stored in maximum register 307.
  • Comparator 308 is provided for comparing output ⁇ M of maximum registger 307 and output ⁇ o of last maximum register 302. If ⁇ o - ⁇ M ⁇ ⁇ ( ⁇ is a threshold value), comparator 308 outputs a logic "1" on line 323. If ⁇ o - ⁇ M ⁇ ⁇ , comparator 308 outputs a logic "O".
  • loop control circuit 400 When the signal on line 323 is "O", loop control circuit 400 outputs a recalculating signal S3 to address control circuit 300.
  • Circuit 300 provides a set signal to last maximum register 302 through line 323 for setting the maximum eigenvalue ⁇ M as the last maximum value ⁇ o .
  • Circuit 300 supplies the same addresses to buffers 301 and 305 through lines 320, 322.
  • Each eigenvalue of buffer 705 is read out and supplied to divider 309.
  • Divider 309 divides each eigenvalue by maximum eigenvalue ⁇ M .
  • the output of divider 309 is stored in buffer 301. After this, recalculation of the eigenvalues is repeated. Recalculation is repeated until ⁇ o - ⁇ M ⁇ ⁇ is satisfied.
  • loop control circuit 400 When the signal on line 323 is "1", loop control circuit 400 outputs address k to eigenvalue memory 11 in order to set the maxiumum eigenvalue ⁇ M as the maximum value of the k-th category. Circuit 400 incremenets k counter 401 which contents indicate the present category being calculated when the contents of the k counter is not K and outputs a repeating signal S4 to address control circuit 200 in Figure 10. Address control circuit 200 controls calculating a new projector Pk and calculating a new matrix QPkQ.
  • Eigenvalue calculating unit 10 also calculates eigenvalues for the new matrix. Therefore, when the contents of k counter 401 becomes K, each maximum eigenvalue for each category is stored in eigenvalue memory 11. At that time, circuit 400 provides an end signal to decision unit 12.
  • Figure 12 shows an example of decision unit 12 in Figure 8 which responds to end signal S5.
  • a sorting circuit 500 compares the eigenvalues to each other in the eigenvalue memory 11 and provides the maximum eigenvalue ⁇ M with its address KM to register 501 and the secondary largest eigenvalue ⁇ N with its address KN to register 502. Eigenvalues ⁇ M and ⁇ N are supplied to comparing circuit 503.
  • Comparing circuit 503 provides an accept signal to line 504, where ⁇ M > ⁇ 1 and ⁇ M - ⁇ N > 8 2 , wherein ⁇ 1 and 8 2 are threshold values for decision operation. If ⁇ M ⁇ 8 1 , or A M - AN ⁇ 2 , comparing circuit 503 provides a reject signal on line 505 indicating the apparatus cannot recognize the input pattern.
  • a character code generator 506 is activated comprising a Read Only Memory which generates a character code corresponding to the address data k m representing the category to which the input pattern belongs.
  • the apparatus can recognize more varieties of deformed input patterns than conventional apparatus which utilizes the multiple similarity method.
  • Figure 14 shows another embodiment of matrix generating unit 7 in Figure 8.
  • the unit calculates an X matrix instead of matrix QPQ.
  • Address control circuit 600 supplies addresses to base memories 126, 127, 128, reference pattern memory 8 and matrix memory 9.
  • a first inner product circuit comprising multiplier 601 and accumulator 602 calculates an inner product between one of the reference pattern ⁇ in the reference pattern memory 8 and base ⁇ i selected by selector 605 in response to selection signal A1 from address control circuit 600.
  • a second inner product circuit comprising multiplier 601 and accumulator 602 calculates an inner product between one of the reference patterns 8m and the base ⁇ j selected by selector 606 in response to selection signal AJ from circuit 600.
  • the outputs of these inner product circuits are multiplied by multiplier 607.
  • the output of multiplier 607 is added to the contents of matrix memory 9 by adder 608.
  • the result is written into matrix memory 9.
  • the advantage of the matrix generating unit shown in Figure 4 is that the size of the X matrix is very small compared to the size of matrix QPQ so that the calculation time for generating the matrix and calculating the eigenvalues is reduced.
  • FIG 15 shows another embodiment of basis generating unit 5 in Figure 8.
  • each base is derived by a direct masking operation of the input vector.
  • mask memories 108' and 109' which store differential blurring masks.
  • the mask in memory 108' is derived by partial-differentiation of the Gaussian function to the x-direction such as shown in Figure 16A.
  • the mask in memory 109' is derived by partial-differentiation of the Gaussian function to the y-direction such as shown in Figure 16B.
  • the first blurred pattern g is obtained by sum-of-product circuit 113 according to a calculation between input vector f and mask pattern data stored in mask memory 107.
  • the second blurred pattern gx and the third blurred pattern gy are obtained by the same operation and they are stored in blurred pattern memories 111' and 112', respectively.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Data Mining & Analysis (AREA)
  • Multimedia (AREA)
  • Evolutionary Computation (AREA)
  • Databases & Information Systems (AREA)
  • Software Systems (AREA)
  • Medical Informatics (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Image Analysis (AREA)
  • Character Discrimination (AREA)

Claims (6)

1. Zeichenerkennungsgerät zur Bestimmung der Kategorie, zu der ein unbekanntes eingegebenes Zwichen gehört, mit einem Speicher (89) zum Speichern einer Vielzahl von Bezugszeichen für jede Kategorie eines vorbestimmten Referenzzeichensatzes, wobei die Referenzzeichen relativ zueinander orthonormalisiert sind, und mit einer Eingabevektor-Erzeugungseinrichtung (1, 2, 4) zum Erzeugen eines Eingabevektors, der Merkmale der unbekannten, eingegebenen Zeichen repräsentiert, gekennzeichnet durch eine basiserzeugende Einrichtung (5), die mit der Eingabevektor-Erzeugungseinrichtung (1, 2, 4) gekoppelt ist zum Deformieren des Eikngabevektors, um eine Vielzahl von Vektoren zu erzeugen und zum Erzeugen von Basisvektoren, die relativ zueinander orthonormalisiert, sind, durch eine matrixerzeugende Einrichtung (7), die mit dem Speicher (8) und der basiserzeugenden Einrichtung (5) gekoppelt ist zur Erzeugung einer Matrix für jede Kategorie des Bezugszeichensatzes, wobei jede Matrix aus einer Vielzahl von Datenelementen gebildet wird, die abgeleitet werden von den Basisvektoren, die durch die basiserzeugende Einrichtung (5) erzeugt werden, sowie von den im Speicher (8) gespeicherten Bezugszeichen, wobei die matrixerzeugende Einrichtung ein projektorerzeugende Einrichtung (204, 205, 206) umfaßt zur Erzeugung eines ersten Projektors, der den durch die Basiserzeugungseinrichtung erzeugten Basen entspricht, und eines zweiten Projektors, der den Bezugszeichen entspricht, und entweder durch eine Multiplikationsrichtung (209), die mit der projektorerzeugenden Einrichtung verbunden ist zum Multiplizieren des ersten und zweiten Projektors, um die Datenelemente für die Matrix zu liefern, oder eine Multiplikationseinrichtung (205) zum Multiplizieren jeder der Basen mit jedem der Bezugszeichen und zum Gewinnen eines Produktes für jede Multiplikation, und eine Summiereinrichtung (206), die mit der Multipliziereinrichtung (205) verbunden ist zum Aufsumieren der Produkte, um die Datenelemente für die Matrix zu erhalten; durch eine Recheneinrichtung (10), die mit der matrixerzeugenden Einrichtung verbunden ist, um den maximalen Eigenwert der Matrix für jede Kategorie des Zeichensatzes zu erhalten; und durch eine Sortiereinrichtung (12), die mit der Recheneinrichtung (10) verbunden ist, zum Selektieren des größen Wertes der maximalen Eigenwerte, wobei der selektierte größte Eigenwert die Kategorie repräsentiert, zu der das eingegebene Zeichen gehört.
2. Zeichenerkennungsgerät nach Anspruch 1, bei dem die Sortiereinrichtung eine Selektionseinrichtung (500, 501, 502) aufweist zum Selektieren des nächsten Eigenwertes unterhalb des ausgewählten größten maximalen Wertes; eine mit der Selektionseinrichtung (500, 501, 502) verbundene Einrichtung (503) vorgesehen ist zum Vergleichen des selektierten maximalen Wertes mit einem vorbestimmten Wert; und eine Einrichtung (503) vorgesehen ist zum Vergleichen des ausgewählten größten Maximalwertes mit dem nächsten ausgewählten Maximalwert, um die Differenz zwischen beiden zu bestimmen.
3. Zeichenerkennungsgerät nach Anspruch 1 oder 2, bei dem die basiserzeugende Einrichtung (5) eine Einrichtung (107, 113) zum Erzeugen eines unscharfen Zeichens aus dem Eingabevektor, eine Einrichtung (108,109,113), die mit der die unscharfen Zeichen erzeugenden Einrichtung verbunden ist, zum Erzeugen einer Vielzahl von Differentialzeichen aus den unscharfen Zeichen, und eine Orthonormalisierungseinrichtung (121, 124, 129, 130) aufweist, die mit der Differentialeinrichtung verbunden ist, zum Gewinnen einer Vielzahl von Basen, die relativ zueiander orthonormalisiert sind.
4. Zeichenerkennungsgerät nach Anspruch 1 oder 2, dadurch gekennzeichnet, daß basiserzeugende Einrichtung (5) folgende Einrichtungen umfaßt: Eine Maskenspeichereinrichtung (107, 108, 109) zum Speichern einer Vielzahl von Maskenzeichen; eine Maskierungseinrichtung (113), die mit der Maskenspeichereinrichtung verbunden ist, zum Ausführen einer Maskenoperation zwischen dem Eingabevektor und jedem der Maskenzeichen, die in der Maskenspeicherinrichtung gespeichert sind, um eine Vielzahl von resultierenden maskierten Zeichen zu erzeugen; und eine orthonormalisierte Einrichtung (121, 124, 129, 130), die mit der Maskierungseinrichtung verbunden ist, zum Orthonormalisieren der resultierenden Maskenzeichen.
5. Zeichenerkennungsgerät nach Anspruch 4, bei dem die Maskenspeichereinrichtung (107, 108, 109) ein unscharfes Maskenzeichen entsprechend einer Gauss-Funktion speichert.
6. Zeichenerkennungsgerät nach einem der Ansprüche 1 bis 5, bei dem die Recheneinrichtung (10) die Eigenwerte der Matrizenn mit Hilfe einer geeigneten Methode berechnet.
EP84306098A 1983-09-07 1984-09-06 Anlage zum erkennen unbekannter MusterE Expired - Lifetime EP0139446B1 (de)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
JP58164853A JPS6057475A (ja) 1983-09-07 1983-09-07 パタ−ン認識方式
JP164853/83 1983-09-07

Publications (3)

Publication Number Publication Date
EP0139446A2 EP0139446A2 (de) 1985-05-02
EP0139446A3 EP0139446A3 (en) 1986-10-08
EP0139446B1 true EP0139446B1 (de) 1990-07-04

Family

ID=15801157

Family Applications (1)

Application Number Title Priority Date Filing Date
EP84306098A Expired - Lifetime EP0139446B1 (de) 1983-09-07 1984-09-06 Anlage zum erkennen unbekannter MusterE

Country Status (4)

Country Link
US (1) US4752957A (de)
EP (1) EP0139446B1 (de)
JP (1) JPS6057475A (de)
DE (1) DE3482637D1 (de)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
DE102012222097A1 (de) 2012-12-03 2014-06-05 Robert Bosch Gmbh Verfahren zum Betreiben eines Kraftstoffeinspritzventils für eine Brennkraftmaschine

Families Citing this family (41)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5142593A (en) * 1986-06-16 1992-08-25 Kabushiki Kaisha Toshiba Apparatus and method for classifying feature data at a high speed
JPS6391699A (ja) * 1986-10-03 1988-04-22 株式会社リコー 音声認識方式
JPH0734228B2 (ja) * 1987-02-23 1995-04-12 株式会社東芝 複合類似度法によるパタ−ン認識装置
US4926488A (en) * 1987-07-09 1990-05-15 International Business Machines Corporation Normalization of speech by adaptive labelling
CA2063723A1 (en) * 1989-07-28 1991-01-29 Stephen J. Guerreri Method and apparatus for language and speaker recognition
US5189727A (en) * 1989-07-28 1993-02-23 Electronic Warfare Associates, Inc. Method and apparatus for language and speaker recognition
EP0441176B1 (de) * 1990-01-25 1994-03-30 Mitsubishi Jidosha Kogyo Kabushiki Kaisha Regelsystem für die Antriebsleistung von Kraftfahrzeugen
JPH0481988A (ja) * 1990-07-24 1992-03-16 Sharp Corp クラスタリング方式
US5164992A (en) * 1990-11-01 1992-11-17 Massachusetts Institute Of Technology Face recognition system
US5167004A (en) * 1991-02-28 1992-11-24 Texas Instruments Incorporated Temporal decorrelation method for robust speaker verification
US5428788A (en) * 1991-05-10 1995-06-27 Siemens Corporate Research, Inc. Feature ratio method for computing software similarity
US5485621A (en) * 1991-05-10 1996-01-16 Siemens Corporate Research, Inc. Interactive method of using a group similarity measure for providing a decision on which groups to combine
US5440742A (en) * 1991-05-10 1995-08-08 Siemens Corporate Research, Inc. Two-neighborhood method for computing similarity between two groups of objects
US5317741A (en) * 1991-05-10 1994-05-31 Siemens Corporate Research, Inc. Computer method for identifying a misclassified software object in a cluster of internally similar software objects
EP0513652A2 (de) * 1991-05-10 1992-11-19 Siemens Aktiengesellschaft Verfahren zur Modellierung einer Ähnlichkeitsfunktion mittels eines Neuronalnetzes
US5438676A (en) * 1991-05-10 1995-08-01 Siemens Corporate Research, Inc. Method for adapting a similarity function for identifying misclassified software objects
US5819016A (en) * 1993-10-05 1998-10-06 Kabushiki Kaisha Toshiba Apparatus for modeling three dimensional information
DE4430348C2 (de) * 1994-08-26 1998-04-02 Rohde & Schwarz Verfahren zum Trennen von Carrier-Signal und Interferer-Signal in einem Funksignal eines Mobilfunknetzes
US5710833A (en) * 1995-04-20 1998-01-20 Massachusetts Institute Of Technology Detection, recognition and coding of complex objects using probabilistic eigenspace analysis
US5920644A (en) * 1996-06-06 1999-07-06 Fujitsu Limited Apparatus and method of recognizing pattern through feature selection by projecting feature vector on partial eigenspace
US6185316B1 (en) 1997-11-12 2001-02-06 Unisys Corporation Self-authentication apparatus and method
JP2001266151A (ja) * 2000-03-17 2001-09-28 Toshiba Corp 個人識別装置および個人識別方法
JP4099981B2 (ja) * 2001-12-04 2008-06-11 日本電気株式会社 画像認識システム、画像認識方法および画像認識プログラム
US20030125946A1 (en) * 2002-01-03 2003-07-03 Wen-Hao Hsu Method and apparatus for recognizing animal species from an animal voice
JP3888676B2 (ja) * 2002-02-25 2007-03-07 株式会社東芝 ３次元物体認識装置及びその方法
US7146037B2 (en) * 2002-07-12 2006-12-05 Winbond Electronics Corp. VLSI neural fuzzy classifier for handwriting recognition
US7124394B1 (en) 2003-04-06 2006-10-17 Luminescent Technologies, Inc. Method for time-evolving rectilinear contours representing photo masks
US7698665B2 (en) * 2003-04-06 2010-04-13 Luminescent Technologies, Inc. Systems, masks, and methods for manufacturable masks using a functional representation of polygon pattern
US8457962B2 (en) * 2005-08-05 2013-06-04 Lawrence P. Jones Remote audio surveillance for detection and analysis of wildlife sounds
JP4588575B2 (ja) * 2005-08-09 2010-12-01 富士フイルム株式会社 デジタル画像の複数対象物検出方法および装置並びにプログラム
WO2007033362A2 (en) * 2005-09-13 2007-03-22 Luminescent Technologies, Inc. Systems, masks, and methods for photolithography
WO2007041600A2 (en) * 2005-10-03 2007-04-12 Luminescent Technologies, Inc. Mask-pattern determination using topology types
US7788627B2 (en) * 2005-10-03 2010-08-31 Luminescent Technologies, Inc. Lithography verification using guard bands
US7793253B2 (en) * 2005-10-04 2010-09-07 Luminescent Technologies, Inc. Mask-patterns including intentional breaks
WO2007044557A2 (en) 2005-10-06 2007-04-19 Luminescent Technologies, Inc. System, masks, and methods for photomasks optimized with approximate and accurate merit functions
JP2007206833A (ja) 2006-01-31 2007-08-16 Toshiba Corp 生体照合方法および生体照合装置
JP2007233873A (ja) * 2006-03-02 2007-09-13 Toshiba Corp パターン認識装置及びその方法
US8116566B2 (en) * 2006-08-28 2012-02-14 Colorado State University Research Foundation Unknown pattern set recognition
US8046200B2 (en) 2006-09-05 2011-10-25 Colorado State University Research Foundation Nonlinear function approximation over high-dimensional domains
US7917540B2 (en) 2007-02-22 2011-03-29 Colorado State University Research Foundation Nonlinear set to set pattern recognition
US9232670B2 (en) 2010-02-02 2016-01-05 Apple Inc. Protection and assembly of outer glass surfaces of an electronic device housing

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPS4912778B1 (de) * 1969-11-05 1974-03-27
US3700815A (en) * 1971-04-20 1972-10-24 Bell Telephone Labor Inc Automatic speaker verification by non-linear time alignment of acoustic parameters
JPS5619656B2 (de) * 1973-08-08 1981-05-08
JPS50155105A (de) * 1974-06-04 1975-12-15
CA1056504A (en) * 1975-04-02 1979-06-12 Visvaldis A. Vitols Keyword detection in continuous speech using continuous asynchronous correlation
JPS5525150A (en) * 1978-08-10 1980-02-22 Nec Corp Pattern recognition unit
US4259661A (en) * 1978-09-01 1981-03-31 Burroughs Corporation Apparatus and method for recognizing a pattern
JPS57178578A (en) * 1981-04-27 1982-11-02 Toshiba Corp Pattern recognition system

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
DE102012222097A1 (de) 2012-12-03 2014-06-05 Robert Bosch Gmbh Verfahren zum Betreiben eines Kraftstoffeinspritzventils für eine Brennkraftmaschine

Also Published As

Publication number Publication date
JPS6057475A (ja) 1985-04-03
EP0139446A3 (en) 1986-10-08
EP0139446A2 (de) 1985-05-02
US4752957A (en) 1988-06-21
DE3482637D1 (de) 1990-08-09

Similar Documents

Publication Publication Date Title
EP0139446B1 (de) Anlage zum erkennen unbekannter MusterE
US4153897A (en) Method and device for detecting the similarity between standard and unknown patterns
US5842194A (en) Method of recognizing images of faces or general images using fuzzy combination of multiple resolutions
US6778685B1 (en) Two-stage local and global fingerprint matching technique for automated fingerprint verification/identification
US4259661A (en) Apparatus and method for recognizing a pattern
US4155072A (en) Character recognition apparatus
US4837842A (en) Character and pattern recognition machine and method
US5901244A (en) Feature extraction system and face image recognition system
US6128410A (en) Pattern matching apparatus and method that considers distance and direction
US5033099A (en) Image recognition system
EP0621556A2 (de) Ein Verfahren zur Kombination der Ergebnisse von mehreren Klassifikatoren
US4630308A (en) Character reader
Samal et al. Human face detection using silhouettes
US6240209B1 (en) Method for deriving character features in a character recognition system
US7415137B2 (en) Image processing method, apparatus and storage medium
US5105470A (en) Method and system for recognizing characters
JPH0696278A (ja) パターン認識方法及び装置
CN111291712B (zh) 基于插值的cn和胶囊网络的森林火灾识别方法及装置
JP3251840B2 (ja) 画像認識装置
JP6075238B2 (ja) 文字認識装置および文字認識方法
EP0435099B1 (de) Kenndatenauswahlverfahren für Datenverarbeitungssystem
US5845020A (en) Character recognizing apparatus
JPH08115387A (ja) パターン認識装置
JP2950023B2 (ja) パターン認識辞書生成装置およびパターン認識装置
JP3421200B2 (ja) 文字認識方法および装置

Legal Events

Date Code Title Description
PUAI Public reference made under article 153(3) epc to a published international application that has entered the european phase

Free format text: ORIGINAL CODE: 0009012

AK Designated contracting states

Designated state(s): DE FR GB IT NL

PUAL Search report despatched

Free format text: ORIGINAL CODE: 0009013

AK Designated contracting states

Kind code of ref document: A3

Designated state(s): DE FR GB IT NL

17P Request for examination filed

Effective date: 19870320

17Q First examination report despatched

Effective date: 19880224

GRAA (expected) grant

Free format text: ORIGINAL CODE: 0009210

AK Designated contracting states

Kind code of ref document: B1

Designated state(s): DE FR GB IT NL

REF Corresponds to:

Ref document number: 3482637

Country of ref document: DE

Date of ref document: 19900809

ET Fr: translation filed
ITF It: translation for a ep patent filed

Owner name: STUDIO CONS. BREVETTUALE S.R.L.

PLBE No opposition filed within time limit

Free format text: ORIGINAL CODE: 0009261

STAA Information on the status of an ep patent application or granted ep patent

Free format text: STATUS: NO OPPOSITION FILED WITHIN TIME LIMIT

26N No opposition filed
ITTA It: last paid annual fee
REG Reference to a national code

Ref country code: GB

Ref legal event code: IF02

PGFP Annual fee paid to national office [announced via postgrant information from national office to epo]

Ref country code: GB

Payment date: 20020904

Year of fee payment: 19

PGFP Annual fee paid to national office [announced via postgrant information from national office to epo]

Ref country code: FR

Payment date: 20020910

Year of fee payment: 19

PGFP Annual fee paid to national office [announced via postgrant information from national office to epo]

Ref country code: DE

Payment date: 20020911

Year of fee payment: 19

PGFP Annual fee paid to national office [announced via postgrant information from national office to epo]

Ref country code: NL

Payment date: 20020930

Year of fee payment: 19

PG25 Lapsed in a contracting state [announced via postgrant information from national office to epo]

Ref country code: GB

Free format text: LAPSE BECAUSE OF NON-PAYMENT OF DUE FEES

Effective date: 20030906

PG25 Lapsed in a contracting state [announced via postgrant information from national office to epo]

Ref country code: NL

Free format text: LAPSE BECAUSE OF NON-PAYMENT OF DUE FEES

Effective date: 20040401

Ref country code: DE

Free format text: LAPSE BECAUSE OF NON-PAYMENT OF DUE FEES

Effective date: 20040401

GBPC Gb: european patent ceased through non-payment of renewal fee

Effective date: 20030906

PG25 Lapsed in a contracting state [announced via postgrant information from national office to epo]

Ref country code: FR

Free format text: LAPSE BECAUSE OF NON-PAYMENT OF DUE FEES

Effective date: 20040528

NLV4 Nl: lapsed or anulled due to non-payment of the annual fee

Effective date: 20040401

REG Reference to a national code

Ref country code: FR

Ref legal event code: ST